Mystery Number Five - Computational Linguistics and Talking Robots - page 97

6.3 Transition from Fixed to Adaptive Behavior
Evolving a software system of fixed behavior into one of adaptive behavior
requires a number of rather obvious extensions:
6.3.1 E XTENSIONS REQUIRED BY AN ADAPTIVE BEHAVIOR AGENT
1. Writable memory
In order to record individual recognition action episodes, the agent's
non-writable memory must be complemented with a writable memory.
2. Decoupling of recognition and action
The agent must be capable of recognition without having to perform the
associated fixed behavior action ( recognition per se ), just as there must be
action triggered by reasoning rather than by a fixed behavior stimulus.
3. Unknowns
The agent must be able to recognize and store unknowns consisting of pre-
viously unencountered constellations of available recognition elements. 12
4. Appraisal
In order to learn from past experiences, the agent must be able to evaluate
the implication of recognitions and the outcome of actions. 13
5. Automatic schema derivation
In order to generalize over similar constellations, the agent must be capable
of automatic schema derivation (Sect. 6.4).
It follows from (1) that fixed behavior and adaptive behavior may be distin-
guished in terms of a non-writable vs. a writable memory. For fixed behavior,
a non-writable memory is sufficient: once a stimulus-response pattern has been
coded, no additional memory is needed for repeating it. An adaptive behavior
agent, however, requires an additional writable memory for recording individ-
ual episodes of recognition and action in the order of their arrival at the now
front . 14
11 In this particular respect, the difference between natural languages boils down to the use of different
placeholders (handles) in the form of language-dependent surfaces.
12 An example from vision is recognition by components (RBC) based on geons, proposed by Biederman
(1987). Cf. L&I'05: “Memory-Based Pattern Completion in Database Semantics.”
13 Ekman 2003, p. 31, speaks of the automatic appraisal mechanism in the “emotion database” of a
cognitive agent. In DBS, appraisal (or evaluation) is integrated into the proplets in the form of values
for an additional appraisal attribute. Thus, the only database needed for appraisal is the Word Bank.
14 In nature, practically all living beings, starting with the protozoa, are capable of some form of asso-
ciative learning (classical conditioning), and must therefore have some writable memory. The guided
pattern method for extending a fixed behavior repertoire (Sect. 6.2) also requires a certain amount of